Plain card for the 2.12.0 release under the mtplx library: memory, speed status and the tensor table

README.md

@@ -2,26 +2,56 @@
 license: other
 license_name: qwen-community-1.0
 license_link: LICENSE
-library_name: mlx
-pipeline_tag: image-text-to-text
+library_name: mtplx
+pipeline_tag: text-generation
 base_model: Qwen/Qwen3.8-Flash-Next
 base_model_relation: quantized
+tags:
+- mtplx
+- mlx
+- apple-silicon
+- macos
+- speculative-decoding
+- multi-token-prediction
+- qwen
+- qwen3.8
+- qwen3.8-flash-next
+- flash-next
+- moe
+- mtp
+- 8-bit
+- vision
+- mac-studio
 ---
 
-# Qwen3.8-Flash-Next-MTPLX-Optimized-Quality
+# Qwen 3.8 Flash-Next Optimized Quality
 
-> **Requires the upcoming MTPLX 2.12.0 release.**
-> The model is published ahead of engine support, which arrives in that release.
+**The 8-bit build of Qwen 3.8 Flash-Next, for Macs with 256 GB or 512 GB. Requires MTPLX 2.12.0 or later.**
 
-Requires MTPLX 2.12.0 or later. Served id: `mtplx-flash-next-optimized-quality`.
-Recipe: `flash-next-optimized-quality`. Source revision: `de4b8e4d43b917e7706784d8bb445c9af86a3540`
-(hub-commit).
+Qwen's 125B-A6B Flash-Next, the Qwen4-generation hybrid mixture of experts with
+Qwen Sparse Attention and a 51B-parameter n-gram table, packed for
+[MTPLX](https://mtplx.com) with its multi-token prediction head. The main model
+and the draft head are 8-bit with group size 64, the structural weights stay in
+BF16, and the n-gram table is 4-bit with group size 32. On a Mac with 128 GB or
+more, [Optimized Speed](https://huggingface.co/Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed) is the
+recommended build.
 
-Verification: **streaming-audited**, mode `streaming`.
-The table below is derived from every written safetensors header, including sidecars.
-BF16 describes storage; the runtime can use a private Q8 hyper-connection copy during decode.
+## Memory
 
-| Tensor class | Stored precision | Payload GB |
+The model weights, the draft head and the vision tower need about 128.5 GiB,
+and the 32 GB n-gram table streams from SSD.
+
+- **128 GB**: Cannot load. The weights, the draft head and the vision tower alone need about 128.5 GiB.
+- **256 GB and 512 GB**: the Macs this pack is for, with about 59.5 GiB left for context and the session cache. The MTPLX app and CLI list it second there, after Optimized Speed.
+
+## Speed
+
+Speed on 256 GB and 512 GB Macs is not measured yet. The 8-bit weights move
+twice the bytes per token of Optimized Speed, so expect slower decoding.
+
+## What is in the pack
+
+| Tensor class | Stored precision | Size (GB) |
 |---|---|---:|
 | attention | Q8/g64 affine; BF16 scales and biases | 0.635044 |
 | embeddings | Q8/g64 affine; BF16 scales and biases | 0.675430 |
@@ -49,30 +79,26 @@ BF16 describes storage; the runtime can use a private Q8 hyper-connection copy d
 | shared expert | Q8/g64 affine; BF16 scales and biases | 0.250675 |
 | vision tower | BF16 | 0.897862 |
 
-Tensor payload: 169.934356 GB (158.264 GiB).
-Files before card/manifest generation: 169.958588 GB.
-`size-checksums.json` records final file sizes and SHA-256, excluding the manifest itself.
-
-- **128 GB**: Cannot load: body, MTP and vision alone are approximately 128.46 GiB.
-- **192 GB**: Stream the n-gram table; use MTPLX_NGRAM_RESIDENT=0. Validate the context workload on this machine.
-- **256 GB**: Fits with the Q4 table resident at 128K (planning estimate; measure the final pack).
+The download is 169.96 GB. `size-checksums.json` lists the
+size and SHA-256 of every other file.
 
-Q8 weights transfer more bytes per parameter than the Speed pack, so bandwidth-limited decode is expected to be slower. Speed is unmeasured until the first run; no speed claim.
-Full-model load verified: **false**.
-The streaming audit checks stored tensors and sampled dequantization against the source.
-It does not prove full-model chat, tool calling, image handling, long-context quality,
-or a Speed-pack performance comparison. Those require a run on a supported Mac.
+## Use it
 
-Recommended memory: **256 GB or 512 GB**. On these Macs, MTPLX's automatic memory
-policy normally keeps the n-gram table in RAM as well as the model weights.
-The SSD-backed table is used when the memory policy calls for it.
+In the Mac app, pick Qwen 3.8 Flash-Next Optimized Quality. From the command line:
 
 ```bash
+pip install mtplx
 mtplx serve --model Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Quality --model-id mtplx-flash-next-optimized-quality
 ```
 
-Base model and weights: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next).
-Qwen Community License, preserved in `LICENSE`.
-Upstream model card preserved as `README-upstream-qwen.md`.
-License and credits carried from [Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed](https://huggingface.co/Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed).
+MTPLX samples at the official Qwen 3.8 settings (temperature 1.0, top-p 0.95,
+top-k 20), and drafts are accepted with exact speculative sampling, so the
+output follows the model's own distribution.
+
+Built with the `flash-next-optimized-quality` recipe from
+[Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) at
+revision `de4b8e4d43b917e7706784d8bb445c9af86a3540`. Qwen Community License, preserved in
+`LICENSE`. The upstream model card is preserved as
+`README-upstream-qwen.md`. License and credits are carried from
+[Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed](https://huggingface.co/Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed).
 Conversion and serving: MTPLX.

size-checksums.json

@@ -9,8 +9,8 @@
       "sha256": "35ca37ccc366f1ba478dab33841a2c0c18ce53fd62f291ca05341f7728b225b2"
     },
     "README.md": {
-      "bytes": 3938,
-      "sha256": "02ac6e8edf6650ac24c5661a20186e95434df2380c09e76c7bf31b141c391633"
+      "bytes": 4101,
+      "sha256": "608a2e2b81ff282a871d1f3bec8e76a8612b30665f2726a8aeada0feb91605e6"
     },
     "chat_template.jinja": {
       "bytes": 8952,
@@ -225,6 +225,6 @@
       "sha256": "ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003"
     }
   },
-  "total_bytes_excluding_manifest": 169958526962,
+  "total_bytes_excluding_manifest": 169958527125,
   "weight_hash_source": "Successful streaming audit from the completed build; immutable weight files reused."
 }